Kinds of evals: code, model and human

The three kindsCopy Anchor Link

Deterministic evals use code: valid JSON, the right field, an exact answer, a regular expression, unit tests passing, a SQL query returning the expected rows. They are fast, free and exact. Model-based evals use a model with a rubric to judge qualities such as faithfulness or tone. They scale, but need calibration against humans. Human evals use people for the final word: expert review, side-by-side comparison, user ratings. They are the most trustworthy and the slowest.

Combining themCopy Anchor Link

Run deterministic checks on every change, model-based checks on every release candidate, and human review on samples and on disagreements. Use humans to label the data that calibrates the model judge, then the judge extends coverage.

MetricsCopy Anchor Link

Choose metrics that match the task: accuracy or F1 for classification, pass rate for code, recall at k for retrieval, win rate for comparisons. Track cost and latency alongside quality. A better score that doubles cost may not be better.

A worked example: one task, three checksCopy Anchor Link

Task: summarize a meeting. Code check: the summary is under 150 words and mentions every attendee's name. Model check: a judge with a rubric asks whether every action item in the notes appears in the summary. Human check: a manager reads ten summaries a week and marks the ones they would send. Cheap checks run on every change, expensive ones on samples.

When a code check liesCopy Anchor Link

Code checks are exact about the rule you wrote, not about what you meant. Suppose the rule is "the answer equals Paris". The reply "The capital is Paris." fails it, though it is correct. The opposite also happens: a summary under 150 words that names every attendee passes, even if it is nonsense.

Symptom: code checks and your own reading disagree on some outputs. Confirm by reading ten failures and ten passes. Fix the rule, for example by checking that the answer contains the key fact, and add a model or human check for what the rule cannot see.

Precision and recall by exampleCopy Anchor Link

Accuracy can hide a useless system. Take 100 emails, 5 of them spam. A filter that flags nothing is right 95 times, so accuracy is 95 percent, yet it catches no spam.

Now a filter flags 10 emails and 4 are truly spam. Precision is the share of flags that were right: 4 of 10, or 40 percent. Recall is the share of real spam it found: 4 of 5, or 80 percent. Which matters more depends on the cost of each mistake. Blocking a real email may hurt more than letting spam through.

Common mistakeCopy Anchor Link

Using a model judge without checking it against people first.